Anticancer protein prediction from sequence with a fine-tuned ESM-2 650M classifier; a BLAST hybrid raises validation AUC to 0.91.
No providers recorded yet. Browse all providers
AntiCP3 predicts whether a protein sequence has anticancer activity. It comes from the Raghava group at IIIT-Delhi, which has maintained anticancer peptide predictors — AntiCP and AntiCP2 — for over a decade. Those tools were trained on short peptides, typically 40 residues or fewer, and do not transfer to full-length proteins: monoclonal antibodies, enzymes, and other protein biologics occupy a different compositional and structural regime. AntiCP3, posted as a bioRxiv preprint in May 2025, is the group's first model built specifically for proteins of 50 to 1,000 residues.
The paper motivates the split with a compositional analysis. Cysteine, phenylalanine, glycine, leucine, and tryptophan are significantly enriched in both anticancer peptides and anticancer proteins, but methionine, glutamine, and tyrosine are enriched only in the proteins, and lysine only in the peptides — evidence that a peptide-trained classifier is scoring the wrong signal when handed a protein.
AntiCP3 is really a family of models, and the distinction matters when choosing which one to use. The artifact published on HuggingFace is a fine-tuned ESM-2 650M sequence classifier that runs on any input sequence with no auxiliary database. The best-performing configuration reported in the paper is not that model alone but a hybrid that adds a BLAST similarity score against the training set to the language model's output; that hybrid is offered through the web server and standalone package rather than as a single downloadable checkpoint.
raghavagps-group/anticp3 repository hosts the fine-tuned ESM-2 t33 650M checkpoint in safetensors format under an MIT license, loadable directly through EsmForSequenceClassification.pip-installable package, a GitHub repository, and the HuggingFace checkpoint.The positive set is 370 natural anticancer proteins retrieved from CancerPPD2 and length-filtered to 50–1,000 residues; CD-HIT at 40% identity reduces these to 205 clusters. Negatives were mined from UniProt by excluding anticancer, antitumor, and antiangiogenic keywords, yielding 558,717 sequences that CD-HIT reduced to 92,388 representatives, from which 370 length-matched sequences were sampled for a balanced dataset. Splits are cluster-aware: 306 sequences per class for training and 64 held out for outer validation, with five-fold cross-validation inside the training fold.
Fine-tuning ESM-2 with a sequence classification head shows a clean scaling trend across checkpoint sizes — AUC 0.79 for the 8M model, 0.82 for 35M, 0.87 for 150M, and 0.90 for the 650M t33 model (MCC 0.63). Feeding embeddings from that fine-tuned model into a separate multilayer perceptron reaches the same AUC of 0.90 with a higher MCC of 0.71 and 85.2% accuracy. The BLAST hybrid at an E-value cutoff of 1e-20 achieves AUC 0.91, MCC 0.63, 84.4% sensitivity, and 78.1% specificity. Deep architectures over raw PSSM matrices — CNN, residual CNN, CNN+BiGRU, CNN+BiLSTM — top out at AUC 0.74, which the authors attribute to the small dataset. A separate experiment that scored proteins by counting AntiCP2-predicted anticancer 10-mers within them reached only AUC 0.58 and was abandoned.
The tool is aimed at researchers screening candidate protein therapeutics: triaging recombinant proteins, natural product isolates, or predicted open reading frames for anticancer activity before committing to cell-based assays. Because it accepts multi-FASTA input through both the standalone and the web server, it fits into proteome-scale prioritization workflows, and the HuggingFace checkpoint allows the pure language model variant to be embedded in a Python pipeline without an external BLAST database.
AntiCP3 is the first predictor dedicated to anticancer proteins rather than peptides, and by releasing a fine-tuned ESM-2 checkpoint under a permissive license the group made a reusable artifact rather than a web form alone. The caveats are substantial and stated plainly in the paper: the entire labelled dataset is 740 sequences, the outer validation set is 128, and the negative class is defined by UniProt keyword exclusion rather than by experimental evidence of inactivity, so reported metrics should be read as an upper bound. The preprint has not yet been peer reviewed, and the top-scoring hybrid configuration depends on BLAST against the small training database, which limits how far it generalizes to sequences with no close homologs.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.